Keratoconus Severity Staging Using Random Forest and Gradient Boosting Ensemble Techniques

Accurate keratoconus (KC) staging is vital for improving patient care and guiding clinical decision making. Choosing the right Machine Learning (ML) algorithm is key to effectively tackling this challenge and ensuring optimal performance. This paper presents a detailed comparison of eight ML algorithms commonly used in KC detection, based on a clinical dataset collected by the authors over the past decade. The study investigates each algorithm's effectiveness in distinguishing KC severity stages. Results showed that ensemble learning algorithms outperformed other approaches, with Random Forest (RF) emerging as the top-performing model, achieving the highest validation accuracy of 98.82%. When tested on unseen data, the RF model maintained a comparable accuracy of 98%, demonstrating its robustness, and generalizability in accurately distinguishing between different stages of KC severity.